From 82cf9005997e73b733105dbb3943d8fbf7cc6156 Mon Sep 17 00:00:00 2001 From: SamAcctX <87765660+SamAcctX@users.noreply.github.com> Date: Fri, 2 Jan 2026 14:24:20 -0600 Subject: [PATCH] Update README.md --- tools/llama-bench/README.md | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/tools/llama-bench/README.md b/tools/llama-bench/README.md index c837bb6d26..879abc4406 100644 --- a/tools/llama-bench/README.md +++ b/tools/llama-bench/README.md @@ -32,6 +32,7 @@ options: -oe, --output-err output format printed to stderr (default: none) --list-devices list available devices and exit -v, --verbose verbose output + -nf, --no-fail continue on failure (default: disabled) --progress print test progress indicators -rpc, --rpc register RPC devices (comma separated) @@ -80,6 +81,10 @@ Each test is repeated the number of times given by `-r`, and the results are ave Using the `-d ` option, each test can be run at a specified context depth, prefilling the KV cache with `` tokens. +Using the `-nf` option, any test param combination that results in a failure will NOT cause the entire set of permuted test scenarios to terminate. Instead, the failing param combination will be logged to STDERR and the execution will then cycle to the next test scenario. This will repeat until all calculated scenarios have been attempted. Test scenarios that successfully execute will log results to STDOUT as usual. +If any permutation resulted in a successful test, the `llama-bench` process will exit with a return code 0 (success). If all permutations failed to execute, the `llama-bench` process will exit with a return code 1 (fail/error). +In order for this new flag to remain backwards-compatible with the `llama-bench` tool's previous behavior for a given invocation, this new mode must be explicitly enabled. + For a description of the other options, see the [completion example](../completion/README.md). > [!NOTE]